Professor Jesus Antonio Izaguirre
Computer Science and Engineering
University of Notre Dame
Notre Dame, IN

Monday, May 2, 2011
3:00 pm (CT)
3269 Beckman Institute

Abstract

The simulation of long timescales of biological relevance in protein function is a grand challenge that requires technological and methodological innovation. I will present a preliminary evaluation of a long time step molecular dynamics (LTMD) method, which allows time steps one or two orders of magnitude larger than current MD algorithms, with real speedups of more than a factor of 20 over MD with constraints. The algorithm coarse grains the dynamics of all atom force fields (such as CHARMM or AMBER) with implicit solvent (such as Generalized Born) using a Mori Zwanzig approach, to produce a Langevin dynamics equation on low frequency mode space coupled to a Brownian dynamics equation on fast frequency mode space. The parameters for the Langevin equation are determined from short (few nanosecond) simulations of unconstrained MD using a novel discrete Mori Zwanzig projection that uses time averages to calculate damping tensors or coefficients. The frequency space is updated using a novel diagonalization method based on partial Hessians augmented with a Krylov subspace method. The diagonalization method is fast and has small memory footprint, and is amenable to numerical computation of Hessians. The latter allows the method to be implemented easily for different force fields. Further improvements for sampling are allowed when the method is combined with replica exchange or with the string method. I will present an evaluation on the method on test peptides and proteins by comparing to MD datasets used as gold standard, and also using Markov State Models to cluster the resulting data. This project is done in collaboration with Dr. Christopher Sweet at the Center for Research Computing at Notre Dame, Prof. Eric Darve at Stanford, and Prof. Vijay S. Pande, Director of Folding@Home at Stanford.